Loading ToolBench

Reading the current corpus…

sepia MCP server · F · ToolBench

sepia

Self-contained MCP server for Sepia — a personal knowledge-graph memory server. Talks to the Sepia REST API (SEPIA_URL + SEPIA_API_KEY); stdio or HTTP transport.

2026-07-28+STREAMABLE_HTTP7 tools

37/100

Definition Quality27

Protocol Readiness55

Supportability43

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence · evidence: Streamable HTTP

Current-spec patterns detected

Summary

Critical failure: Tool definitions are inferred from a SvelteKit dashboard component (tool-breadth.svelte), not from actual MCP server source code. No explicit tool registration, schemas, or parameter definitions are visible in the provided source. The server uses tmcp framework (HTTP transport via @tmcp/transport-http), but the actual tool implementations in packages/sepia-mcp/src/tools/ are not included in the evaluation materials. Tool names exist (manage_memory, manage_entity, etc.) but without access to their input schemas, parameter descriptions, or output structures, scoring must be conservative. Descriptions in the dashboard are generic summaries (e.g., 'Manage memory operations including create, read, update, delete, and ingest, 7 actions total') that do not meet LLM-optimization standards. Per HARD SCORING RULE: tools inferred rather than explicitly defined are capped at 50; since definitions are not visible at all, scores are lower.

Tools (7)

manage_entitywriteauthsource verified25/100

Manage entity operations — 6 actions

manage_memorywriteauthsource verified25/100

Manage memory operations including create, read, update, delete, and ingest (conversation handoff) — 7 actions total

manage_namespacewriteauthsource verified25/100

Manage namespace operations — 4 actions

manage_relationwriteauthsource verified25/100

Manage relation operations — 3 actions

prune_memoriesdestructiveauthsource verified30/100

Prune memories — 1 action

searchread onlyauthsource verified32/100

Search memories — 1 action

traverse_graphread onlyauthsource verified30/100

Traverse knowledge graph — 1 action

Issues (5)

CRITICALtool ↗

No input schemas visible for any tool. Per HARD SCORING RULE, schema score must be 0 when not visible. Cannot verify parameter types, constraints, enums, or descriptions.

manage_memorymanage_entitymanage_namespacemanage_relationprune_memoriessearchtraverse_graph

CRITICALtool ↗

Tool definitions inferred from dashboard component, not from actual MCP server registration code. Per HARD SCORING RULE, inferred tools capped at 50; since source is not visible, scores are lower. Need to see packages/sepia-mcp/src/tools/*.ts implementations.

manage_memorymanage_entitymanage_namespacemanage_relationprune_memoriessearchtraverse_graph

CRITICALtool-description ↗

Descriptions are generic summaries under 50 characters (e.g., 'Manage memory operations, 7 actions total'). Per HARD SCORING RULE, descriptions under 20 chars score 0-20. These lack WHAT the tool does, WHEN to use it, and prerequisites. LLM cannot determine when to select manage_memory vs manage_entity.

Recommendations

Provide the actual tool implementation files (packages/sepia-mcp/src/tools/*.ts) showing explicit tool registration with tmcp, including input schemas, parameter descriptions, and output types.
Expand tool descriptions to 50 - 200 characters following LLM-optimization guidelines: state WHAT (action), WHEN (use case), and any prerequisites. Example: 'Create a new memory entry in the knowledge graph. Use this to record facts, relationships, or conversation context. Requires a namespace and entity type.'
Add comprehensive parameter documentation: type, description, constraints (enum, min/max, pattern), and whether required. Example: 'namespace (string, required): The knowledge graph namespace (2-50 alphanumeric chars, e.g., "personal", "work").'
Document output schemas for each tool showing returned fields, types, and examples. Include IDs and references needed for chaining (e.g., if manage_memory returns memory_id, ensure downstream tools accept it).
Implement error handling that guides LLM recovery: categorize errors as retryable/user-fixable/fatal, include the invalid value and constraint violated, and suggest next steps. Example: 'Invalid namespace: got "my-ns", must be 2-50 alphanumeric chars. Try: create_namespace("my_ns") first.'

Score history

Overall score trend

First recorded score · v2 rubric

37/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	F	37	2026-07-28+	v2